ZFS file system id is not persistent

[imp]

When moving zpool between different hosts the file system id of a given filesystem is not persistent. If this procedure is used for NFS servers failover NFS clients have little chance to survive as their NFS handles become stale.

Apparently commit SHA: 53cf50e introduced the next code in fs_statvfs()

        statp->f_fsid.val[0] = dentry->d_sb->s_dev;
        statp->f_fsid.val[1] = 0;

d_sb->s_dev may be different on different hosts, since it comes from (if I am not mistaken) mount_nodev()/get_sb_nodev().

[imp]

I have experimented with a quick patch that makes f_fsid.val to be zfs dataset guid and it appears to work quite well.

[behlendorf]

Yes, I can see how that would be a problem. d_sb->s_dev may differ between hosts. It looks like internally OpenSolaris is also using the guid for the fsid so that's a good choice.

usr/src/uts/common/fs/zfs/zfs_vfsops.c:zfs_domount()
...

    /*
     * The fsid is 64 bits, composed of an 8-bit fs type, which
     * separates our fsid from any other filesystem types, and a
     * 56-bit objset unique ID.  The objset unique ID is unique to
     * all objsets open on this system, provided by unique_create().
     * The 8-bit fs type must be put in the low bits of fsid[1]
     * because that's where other Solaris filesystems put it.
     */
    fsid_guid = dmu_objset_fsid_guid(zfsvfs->z_os);
    ASSERT((fsid_guid & ~((1ULL<<56)-1)) == 0);
    vfsp->vfs_fsid.val[0] = fsid_guid;
    vfsp->vfs_fsid.val[1] = ((fsid_guid>>32) << 8) |
        zfsfstype & 0xFF;

My only issue with the patch is that we should probably ensure a known byte ordering so this doesn't change between little/big endian systems. The ext4_statfs() function has one reasonable example of this.

[imp]

I've spent some time looking through the code that uses f_fsid.val and it is my understanding that it is used as a value, rather than as a byte array. Since GUID itself is endian-neutral it wan't change between little/big endian systems. Unfortunately I lack access to any Linux capable big endian machine in order ot prove it.

Illumos/OpenSolaris, by the way does the same - the code doesn't do anything special for little/big endian systems.

If, however, I have missed something obvious I'll be glad to stand corrected.